Absolute and Relative Depth-Induced Network for RGB-D Salient Object Detection

Round 1

Reviewer 1 Report

The authors propose using a Depth-Induced Network (DIN) for RGB-D salient object detection to take full advantage of both absolution and relative depth information and further enforce the in-depth fusion of the RGB-D cross-modalities. The proposed DIN is a lightweight network, and the model size is much smaller than that of state-of-the-art algorithms. Extensive experiments on six challenging benchmarks show that the proposed method is feasible. 

Overall, the submission is written well and innovative. However, there are some minor comments on the improvements.

1. The baseline network in Fig. 1(d) is not described or no reference given.

2. Fig. 2 should appear in Sec. 4.2 since all the evaluated models are not mentioned and the performance metrics are not given. The figure legend should also be modified for there are no curves in the figure.

3. Sec. 2 heading “Related work” è “Related Works”.

4. The parentheses are not equal in Eq. (7).

5. “Tab.” need not be shortened in text.

Author Response

Please see the attachment.

Reviewer 2 Report

It's a very good manuscipt.

What about if the depth maps are not available or they are not very accurate? Few comments are needed that describe/compare RGB-D and RGB methods. In turn, authors may compare it with such a work 'Pyramidal attention for saliency detection, CVPR2022'.

Few notations can be revised to be consistent. Like, in Fig. 3 and Eq. 1

[.;.]  --> c

Up --> U

‘MLP is mentioned in Eq. 5 & 6, but it’s architecture is not described anywhere.’

Author Response

Please see the attachment.

Author Response File: Author Response.pdf